COMPMID-1303: CLDepthConvert : Add support for FP32 -> FP16 and FP16 -> FP32 + validate() function

Change-Id: I6808de0254a7c4bca440322cc14b795b3b32465b
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/142427
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

8 files changed, 244 insertions, 70 deletions

diff --git a/arm_compute/core/CL/kernels/CLDepthConvertLayerKernel.h b/arm_compute/core/CL/kernels/CLDepthConvertLayerKernel.h
index 7e795c607a..b70a6a993a 100644
--- a/arm_compute/core/CL/kernels/CLDepthConvertLayerKernel.h
+++ b/arm_compute/core/CL/kernels/CLDepthConvertLayerKernel.h
@@ -48,13 +48,25 @@ public:
      *   - S16 -> U8, U32, S32
      *   - U32 -> U8, U16, S16
      *   - S32 -> U8, U16, S16
+     *   - F16 -> F32
+     *   - F32 -> F16
      *
-     * @param[in]  input  The input tensor to convert. Data types supported: U8/U16/S16/U32/S32/F32.
-     * @param[out] output The output tensor. Data types supported: U8/U16/S16/U32/S32/F32.
+     * @param[in]  input  The input tensor to convert. Data types supported: U8/U16/S16/U32/S32/F16/F32.
+     * @param[out] output The output tensor. Data types supported: U8/U16/S16/U32/S32/F16/F32.
      * @param[in]  policy Conversion policy
      * @param[in]  shift  Value for down/up conversions. Must be 0 <= shift < 8.
      */
     void configure(const ICLTensor *input, ICLTensor *output, ConvertPolicy policy, uint32_t shift);
+    /** Static function to check if given info will lead to a valid configuration of @ref CLDepthConvertLayerKernel
+     *
+     * @param[in] input  Source tensor info. Data types supported: U8/U16/S16/U32/S32/F16/F32.
+     * @param[in] output Destination tensor info. Data type supported: U8/U16/S16/U32/S32/F16/F32.
+     * @param[in] policy Conversion policy
+     * @param[in] shift  Value for down/up conversions. Must be 0 <= shift < 8.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, ConvertPolicy policy, uint32_t shift);
 };
 } // namespace arm_compute
 #endif /*__ARM_COMPUTE_CLDEPTHCONVERTKERNEL_H__ */
diff --git a/arm_compute/runtime/CL/functions/CLDepthConvertLayer.h b/arm_compute/runtime/CL/functions/CLDepthConvertLayer.h
index 00fa0a686d..40ae907805 100644
--- a/arm_compute/runtime/CL/functions/CLDepthConvertLayer.h
+++ b/arm_compute/runtime/CL/functions/CLDepthConvertLayer.h
@@ -48,13 +48,25 @@ public:
      *   - S16 -> U8, U32, S32
      *   - U32 -> U8, U16, S16
      *   - S32 -> U8, U16, S16
+     *   - F16 -> F32
+     *   - F32 -> F16
      *
-     * @param[in]  input  The input tensor to convert. Data types supported: U8/U16/S16/Q16/U32/S32/F32.
-     * @param[out] output The output tensor. Data types supported: U8/U16/S16/U32/S32/F32.
+     * @param[in]  input  The input tensor to convert. Data types supported: U8/U16/S16/U32/S32/F16/F32.
+     * @param[out] output The output tensor. Data types supported: U8/U16/S16/U32/S32/F16/F32.
      * @param[in]  policy Conversion policy.
      * @param[in]  shift  Value for down/up conversions. Must be 0 <= shift < 8.
      */
     void configure(const ICLTensor *input, ICLTensor *output, ConvertPolicy policy, uint32_t shift);
+    /** Static function to check if given info will lead to a valid configuration of @ref CLDepthConvertLayer
+     *
+     * @param[in] input  Source tensor info. Data types supported: U8/U16/S16/U32/S32/F16/F32.
+     * @param[in] output Destination tensor info. Data type supported: U8/U16/S16/U32/S32/F16/F32.
+     * @param[in] policy Conversion policy.
+     * @param[in] shift  Value for down/up conversions. Must be 0 <= shift < 8.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *output, ConvertPolicy policy, uint32_t shift);
 };
 }
 #endif /*__ARM_COMPUTE_CLDEPTHCONVERT_H__*/
diff --git a/src/core/CL/cl_kernels/depth_convert.cl b/src/core/CL/cl_kernels/depth_convert.cl
index 01491ec1b7..611449e614 100644
--- a/src/core/CL/cl_kernels/depth_convert.cl
+++ b/src/core/CL/cl_kernels/depth_convert.cl
@@ -24,8 +24,13 @@
 #include "helpers.h"
 
 #ifdef SATURATE
+#if defined(IS_DATA_TYPE_FLOAT)
+#define CONVERT_RTE(x, type) (convert_##type##_rte((x)))
+#define CONVERT_DOWN(x, type) CONVERT_RTE(x, type)
+#else /* defined(IS_DATA_TYPE_FLOAT) */
 #define CONVERT_DOWN(x, type) CONVERT_SAT(x, type)
-#else /* SATURATE */
+#endif /* defined(IS_DATA_TYPE_FLOAT) */
+#else  /* SATURATE */
 #define CONVERT_DOWN(x, type) CONVERT(x, type)
 #endif /* SATURATE */
 
@@ -36,13 +41,13 @@
  * @attention The input and output data_types need to be passed at compile time using -DDATA_TYPE_IN and -DDATA_TYPE_OUT:
  * e.g. -DDATA_TYPE_IN=uchar -DDATA_TYPE_OUT=short
  *
- * @param[in]  in_ptr                            Pointer to the source image. Supported data types: U8, U16, S16, U32, S32, F16, F32
+ * @param[in]  in_ptr                            Pointer to the source image. Supported data types: U8/U16/S16/U32/S32/F16/F32
  * @param[in]  in_stride_x                       Stride of the source image in X dimension (in bytes)
  * @param[in]  in_step_x                         in_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  in_stride_y                       Stride of the source image in Y dimension (in bytes)
  * @param[in]  in_step_y                         in_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  in_offset_first_element_in_bytes  The offset of the first element in the source image
- * @param[out] out_ptr                           Pointer to the destination image. Supported data types: U8, U16, S16, U32, S32
+ * @param[out] out_ptr                           Pointer to the destination image. Supported data types: U8/U16/S16/U32/S32/F16/F32
  * @param[in]  out_stride_x                      Stride of the destination image in X dimension (in bytes)
  * @param[in]  out_step_x                        out_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  out_stride_y                      Stride of the destination image in Y dimension (in bytes)
@@ -63,7 +68,12 @@ __kernel void convert_depth_down(
     VEC_DATA_TYPE(DATA_TYPE_IN, 16)
     in_data = vload16(0, (__global DATA_TYPE_IN *)in.ptr);
 
+#if defined(IS_DATA_TYPE_FLOAT)
+    const DATA_TYPE_IN scale = (DATA_TYPE_IN)(1 << shift);
+    vstore16(CONVERT_DOWN(in_data / scale, VEC_DATA_TYPE(DATA_TYPE_OUT, 16)), 0, (__global DATA_TYPE_OUT *)out.ptr);
+#else  /* defined(IS_DATA_TYPE_FLOAT) */
     vstore16(CONVERT_DOWN(in_data >> shift, VEC_DATA_TYPE(DATA_TYPE_OUT, 16)), 0, (__global DATA_TYPE_OUT *)out.ptr);
+#endif /* defined(IS_DATA_TYPE_FLOAT) */
 }
 
 /** This function performs a up-scaling depth conversion.
@@ -71,13 +81,13 @@ __kernel void convert_depth_down(
  * @attention The input and output data_types need to be passed at compile time using -DDATA_TYPE_IN and -DDATA_TYPE_OUT:
  * e.g. -DDATA_TYPE_IN=uchar -DDATA_TYPE_OUT=short
  *
- * @param[in]  in_ptr                            Pointer to the source image. Supported data types: U8, U16, S16, U32 or S32
+ * @param[in]  in_ptr                            Pointer to the source image. Supported data types: U8/U16/S16/U32/S32/F16/F32
  * @param[in]  in_stride_x                       Stride of the source image in X dimension (in bytes)
  * @param[in]  in_step_x                         in_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  in_stride_y                       Stride of the source image in Y dimension (in bytes)
  * @param[in]  in_step_y                         in_stride_y * number of elements along Y processed per workitem(in bytes)
  * @param[in]  in_offset_first_element_in_bytes  The offset of the first element in the source image
- * @param[out] out_ptr                           Pointer to the destination image. Supported data types: U8, U16, S16, U32, S32, F16 or F32
+ * @param[out] out_ptr                           Pointer to the destination image. Supported data types: U8/U16/S16/U32/S32/F16/F32
  * @param[in]  out_stride_x                      Stride of the destination image in X dimension (in bytes)
  * @param[in]  out_step_x                        out_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  out_stride_y                      Stride of the destination image in Y dimension (in bytes)
@@ -98,5 +108,10 @@ __kernel void convert_depth_up(
     VEC_DATA_TYPE(DATA_TYPE_IN, 16)
     in_data = vload16(0, (__global DATA_TYPE_IN *)in.ptr);
 
+#if defined(IS_DATA_TYPE_FLOAT)
+    const DATA_TYPE_OUT scale = (DATA_TYPE_OUT)(1 << shift);
+    vstore16(CONVERT_UP(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, 16)) * scale, 0, (__global DATA_TYPE_OUT *)out.ptr);
+#else  /* defined(IS_DATA_TYPE_FLOAT) */
     vstore16(CONVERT_UP(in_data, VEC_DATA_TYPE(DATA_TYPE_OUT, 16)) << shift, 0, (__global DATA_TYPE_OUT *)out.ptr);
+#endif /* defined(IS_DATA_TYPE_FLOAT) */
 }
diff --git a/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp b/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp
index 2f5b2466b1..ffbd295646 100644
--- a/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp
+++ b/src/core/CL/kernels/CLDepthConvertLayerKernel.cpp
@@ -25,6 +25,7 @@
 
 #include "arm_compute/core/CL/CLHelpers.h"
 #include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/CLValidate.h"
 #include "arm_compute/core/CL/ICLTensor.h"
 #include "arm_compute/core/CL/OpenCL.h"
 #include "arm_compute/core/Error.h"
@@ -38,64 +39,83 @@
 
 using namespace arm_compute;
 
-void CLDepthConvertLayerKernel::configure(const ICLTensor *input, ICLTensor *output, ConvertPolicy policy, uint32_t shift)
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, ConvertPolicy policy, uint32_t shift)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S16,
-                                                  DataType::U16, DataType::U32, DataType::S32);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::S16,
-                                                  DataType::U16, DataType::U32, DataType::S32, DataType::F32);
-    ARM_COMPUTE_ERROR_ON(input == output);
-    ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == output->info()->data_type(), "Input and output data types must be different");
-    ARM_COMPUTE_ERROR_ON(shift >= 8);
+    ARM_COMPUTE_UNUSED(policy);
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
+    ARM_COMPUTE_RETURN_ERROR_ON(input == output);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S16,
+                                                         DataType::U16, DataType::U32, DataType::S32,
+                                                         DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8, DataType::S16,
+                                                         DataType::U16, DataType::U32, DataType::S32,
+                                                         DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == output->data_type(), "Input and output data types must be different");
+    ARM_COMPUTE_RETURN_ERROR_ON(shift >= 8);
 
     // Check if convertion is supported
-    ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::U8 && (output->info()->data_type() != DataType::U16 && output->info()->data_type() != DataType::S16
-                                                                            && output->info()->data_type() != DataType::U32 && output->info()->data_type() != DataType::S32),
-                             "Only data types supported [in] U8 -> [out] U16, S16, U32, S32");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::U8 && (output->data_type() != DataType::U16 && output->data_type() != DataType::S16
+                                                                           && output->data_type() != DataType::U32 && output->data_type() != DataType::S32),
+                                    "Only data types supported [in] U8 -> [out] U16, S16, U32, S32");
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::U16 && (output->data_type() != DataType::U8 && output->data_type() != DataType::U32
+                                                                            && output->data_type() != DataType::S32),
+                                    "Only data types supported [in] U16 ->  [out] U8, U32, S32");
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::S16 && (output->data_type() != DataType::U8 && output->data_type() != DataType::U32
+                                                                            && output->data_type() != DataType::S32),
+                                    "Only data types supported [in] S16 ->  [out] U8, U32, S32");
+
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::U32 && (output->data_type() != DataType::U8 && output->data_type() != DataType::U16
+                                                                            && output->data_type() != DataType::S16),
+                                    "Only data types supported [in] U32 ->  [out] U8, U16, S16");
 
-    ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::U16 && (output->info()->data_type() != DataType::U8 && output->info()->data_type() != DataType::U32
-                                                                             && output->info()->data_type() != DataType::S32),
-                             "Only data types supported [in] U16 ->  [out] U8, U32, S32");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::S32 && (output->data_type() != DataType::U8 && output->data_type() != DataType::U16
+                                                                            && output->data_type() != DataType::S16),
+                                    "Only data types supported [in] S32 ->  [out] U8, U16, S16");
 
-    ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::S16 && (output->info()->data_type() != DataType::U8 && output->info()->data_type() != DataType::U32
-                                                                             && output->info()->data_type() != DataType::S32),
-                             "Only data types supported [in] S16 ->  [out] U8, U32, S32");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::F16 && output->data_type() != DataType::F32,
+                                    "Only data types supported [in] F16 ->  [out] F32");
 
-    ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::U32 && (output->info()->data_type() != DataType::U8 && output->info()->data_type() != DataType::U16
-                                                                             && output->info()->data_type() != DataType::S16),
-                             "Only data types supported [in] U32 ->  [out] U8, U16, S16");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::F32 && output->data_type() != DataType::F16,
+                                    "Only data types supported [in] F32 ->  [out] F16");
+
+    // Validate in case of configured output
+    if(output->total_size() > 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
+    }
+
+    return Status{};
+}
+} // namespace
 
-    ARM_COMPUTE_ERROR_ON_MSG(input->info()->data_type() == DataType::S32 && (output->info()->data_type() != DataType::U8 && output->info()->data_type() != DataType::U16
-                                                                             && output->info()->data_type() != DataType::S16),
-                             "Only data types supported [in] S32 ->  [out] U8, U16, S16");
+void CLDepthConvertLayerKernel::configure(const ICLTensor *input, ICLTensor *output, ConvertPolicy policy, uint32_t shift)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
 
     // Auto initialize output shape if not initialized (We can only auto-configure the shape, datatype must be given)
     set_shape_if_empty(*output->info(), input->info()->tensor_shape());
 
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output);
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), policy, shift));
 
     // Get data sizes
     const size_t input_size  = data_size_from_type(input->info()->data_type());
     const size_t output_size = data_size_from_type(output->info()->data_type());
 
-    // Construct kernel name and build options
-    std::string           kernel_name = "convert_depth";
-    std::set<std::string> build_opts;
-    if(input_size > output_size)
-    {
-        kernel_name += "_down";
-        // Down conversions from float always SATURATE as out-of-bounds conversion from float->integer is implementation defined
-        build_opts.insert(((policy == ConvertPolicy::WRAP) && !is_data_type_float(input->info()->data_type())) ? "-DWRAP" : "-DSATURATE");
-    }
-    else
-    {
-        kernel_name += "_up";
-    }
-    build_opts.emplace("-DDATA_TYPE_IN=" + get_cl_type_from_data_type(input->info()->data_type()));
-    build_opts.emplace("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type()));
+    // Set build options
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DDATA_TYPE_IN=" + get_cl_type_from_data_type(input->info()->data_type()));
+    build_opts.add_option("-DDATA_TYPE_OUT=" + get_cl_type_from_data_type(output->info()->data_type()));
+    // Down conversions from float always SATURATE as out-of-bounds conversion from float->integer is implementation defined
+    build_opts.add_option_if(input_size > output_size, ((policy == ConvertPolicy::WRAP) && !is_data_type_float(input->info()->data_type())) ? "-DWRAP" : "-DSATURATE");
+    build_opts.add_option_if(is_data_type_float(input->info()->data_type()), "-DIS_DATA_TYPE_FLOAT");
 
     // Create kernel
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts));
+    const std::string kernel_name = (input_size > output_size) ? "convert_depth_down" : "convert_depth_up";
+    _kernel                       = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
 
     // Set shift arg
     unsigned int idx = 2 * num_arguments_per_2D_tensor(); //Skip the input and output parameters
@@ -105,3 +125,10 @@ void CLDepthConvertLayerKernel::configure(const ICLTensor *input, ICLTensor *out
     constexpr unsigned int num_elems_processed_per_iteration = 16;
     ICLSimple2DKernel::configure(input, output, num_elems_processed_per_iteration);
 }
+
+Status CLDepthConvertLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, ConvertPolicy policy, uint32_t shift)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, policy, shift));
+
+    return Status{};
+}
diff --git a/src/runtime/CL/functions/CLDepthConvertLayer.cpp b/src/runtime/CL/functions/CLDepthConvertLayer.cpp
index b448465909..2e52e8aadc 100644
--- a/src/runtime/CL/functions/CLDepthConvertLayer.cpp
+++ b/src/runtime/CL/functions/CLDepthConvertLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2017 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -36,3 +36,8 @@ void CLDepthConvertLayer::configure(const ICLTensor *input, ICLTensor *output, C
     k->configure(input, output, policy, shift);
     _kernel = std::move(k);
 }
+
+Status CLDepthConvertLayer::validate(const ITensorInfo *input, const ITensorInfo *output, ConvertPolicy policy, uint32_t shift)
+{
+    return CLDepthConvertLayerKernel::validate(input, output, policy, shift);
+}
diff --git a/tests/validation/CL/DepthConvertLayer.cpp b/tests/validation/CL/DepthConvertLayer.cpp
index ed1f54ca6e..fe46313568 100644
--- a/tests/validation/CL/DepthConvertLayer.cpp
+++ b/tests/validation/CL/DepthConvertLayer.cpp
@@ -44,15 +44,16 @@ namespace validation
 namespace
 {
 /** Input data sets **/
-const auto DepthConvertLayerU8toU16Dataset             = combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::U16));
-const auto DepthConvertLayerU8toS16Dataset             = combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::S16));
-const auto DepthConvertLayerU8toS32Dataset             = combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::S32));
-const auto DepthConvertLayerU16toU8Dataset             = combine(framework::dataset::make("DataType", DataType::U16), framework::dataset::make("DataType", DataType::U8));
-const auto DepthConvertLayerU16toU32Dataset            = combine(framework::dataset::make("DataType", DataType::U16), framework::dataset::make("DataType", DataType::U32));
-const auto DepthConvertLayerS16toU8Dataset             = combine(framework::dataset::make("DataType", DataType::S16), framework::dataset::make("DataType", DataType::U8));
-const auto DepthConvertLayerS16toS32Dataset            = combine(framework::dataset::make("DataType", DataType::S16), framework::dataset::make("DataType", DataType::S32));
-const auto DepthConvertLayerShiftDataset               = framework::dataset::make("Shift", 0, 7);
-const auto DepthConvertLayerFixedPointQuantizedDataset = framework::dataset::make("FractionalBits", 1, 7);
+const auto DepthConvertLayerU8toU16Dataset  = combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::U16));
+const auto DepthConvertLayerU8toS16Dataset  = combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::S16));
+const auto DepthConvertLayerU8toS32Dataset  = combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::S32));
+const auto DepthConvertLayerU16toU8Dataset  = combine(framework::dataset::make("DataType", DataType::U16), framework::dataset::make("DataType", DataType::U8));
+const auto DepthConvertLayerU16toU32Dataset = combine(framework::dataset::make("DataType", DataType::U16), framework::dataset::make("DataType", DataType::U32));
+const auto DepthConvertLayerS16toU8Dataset  = combine(framework::dataset::make("DataType", DataType::S16), framework::dataset::make("DataType", DataType::U8));
+const auto DepthConvertLayerS16toS32Dataset = combine(framework::dataset::make("DataType", DataType::S16), framework::dataset::make("DataType", DataType::S32));
+const auto DepthConvertLayerF16toF32Dataset = combine(framework::dataset::make("DataType", DataType::F16), framework::dataset::make("DataType", DataType::F32));
+const auto DepthConvertLayerF32toF16Dataset = combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::F16));
+const auto DepthConvertLayerShiftDataset    = framework::dataset::make("Shift", 0, 7);
 } // namespace
 
 TEST_SUITE(CL)
@@ -67,6 +68,10 @@ template <typename T>
 using CLDepthConvertLayerToU8Fixture = DepthConvertLayerValidationFixture<CLTensor, CLAccessor, CLDepthConvertLayer, T, uint8_t>;
 template <typename T>
 using CLDepthConvertLayerToU32Fixture = DepthConvertLayerValidationFixture<CLTensor, CLAccessor, CLDepthConvertLayer, T, uint32_t>;
+template <typename T>
+using CLDepthConvertLayerToF16Fixture = DepthConvertLayerValidationFixture<CLTensor, CLAccessor, CLDepthConvertLayer, T, half>;
+template <typename T>
+using CLDepthConvertLayerToF32Fixture = DepthConvertLayerValidationFixture<CLTensor, CLAccessor, CLDepthConvertLayer, T, float>;
 
 TEST_SUITE(U8_to_U16)
 DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(framework::dataset::concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
@@ -343,6 +348,84 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthConvertLayerToS32Fixture<int16_t>, frame
 }
 TEST_SUITE_END()
 
+TEST_SUITE(F16_to_F32)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(framework::dataset::concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
+                                                                   DepthConvertLayerShiftDataset),
+               shape, policy, shift)
+{
+    // Create tensors
+    CLTensor src = create_tensor<CLTensor>(shape, DataType::F16, 1);
+    CLTensor dst = create_tensor<CLTensor>(shape, DataType::F32, 1);
+
+    // Create and Configure function
+    CLDepthConvertLayer depth_convert;
+    depth_convert.configure(&src, &dst, policy, shift);
+
+    // Validate valid region
+    const ValidRegion valid_region = shape_to_valid_region(shape);
+    validate(dst.info()->valid_region(), valid_region);
+
+    // Validate padding
+    const PaddingSize padding = PaddingCalculator(shape.x(), 16).required_padding();
+    validate(src.info()->padding(), padding);
+    validate(dst.info()->padding(), padding);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthConvertLayerToF32Fixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), DepthConvertLayerF16toF32Dataset),
+                                                                                                                   framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
+                                                                                                                   DepthConvertLayerShiftDataset))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthConvertLayerToF32Fixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), DepthConvertLayerF16toF32Dataset),
+                                                                                                                 framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
+                                                                                                                 DepthConvertLayerShiftDataset))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+TEST_SUITE_END()
+
+TEST_SUITE(F32_to_F16)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(framework::dataset::concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
+                                                                   DepthConvertLayerShiftDataset),
+               shape, policy, shift)
+{
+    // Create tensors
+    CLTensor src = create_tensor<CLTensor>(shape, DataType::F32, 1);
+    CLTensor dst = create_tensor<CLTensor>(shape, DataType::F16, 1);
+
+    // Create and Configure function
+    CLDepthConvertLayer depth_convert;
+    depth_convert.configure(&src, &dst, policy, shift);
+
+    // Validate valid region
+    const ValidRegion valid_region = shape_to_valid_region(shape);
+    validate(dst.info()->valid_region(), valid_region);
+
+    // Validate padding
+    const PaddingSize padding = PaddingCalculator(shape.x(), 16).required_padding();
+    validate(src.info()->padding(), padding);
+    validate(dst.info()->padding(), padding);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthConvertLayerToF16Fixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), DepthConvertLayerF32toF16Dataset),
+                                                                                                                    framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
+                                                                                                                    DepthConvertLayerShiftDataset))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthConvertLayerToF16Fixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), DepthConvertLayerF32toF16Dataset),
+                                                                                                                  framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
+                                                                                                                  DepthConvertLayerShiftDataset))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+TEST_SUITE_END()
+
 TEST_SUITE_END()
 TEST_SUITE_END()
 } // namespace validation
diff --git a/tests/validation/reference/DepthConvertLayer.cpp b/tests/validation/reference/DepthConvertLayer.cpp
index 6f90963360..fd2e0ae378 100644
--- a/tests/validation/reference/DepthConvertLayer.cpp
+++ b/tests/validation/reference/DepthConvertLayer.cpp
@@ -60,6 +60,33 @@ SimpleTensor<T2> depth_convert(const SimpleTensor<T1> &src, DataType dt_out, Con
     return result;
 }
 
+template < typename T1, typename T2, typename std::enable_if < is_floating_point<T1>::value &&is_floating_point<T2>::value &&!std::is_same<T1, T2>::value, int >::type >
+SimpleTensor<T2> depth_convert(const SimpleTensor<T1> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift)
+{
+    SimpleTensor<T2> result(src.shape(), dt_out);
+
+    const uint32_t scale = 1 << shift;
+
+    // Up-casting
+    if(src.data_type() <= dt_out)
+    {
+        for(int i = 0; i < src.num_elements(); ++i)
+        {
+            result[i] = src[i] * static_cast<T2>(scale);
+        }
+    }
+    // Down-casting
+    else
+    {
+        for(int i = 0; i < src.num_elements(); ++i)
+        {
+            T1 val    = src[i] / static_cast<T1>(scale);
+            result[i] = (policy == ConvertPolicy::SATURATE) ? saturate_cast<T2>(val) : static_cast<T2>(val);
+        }
+    }
+    return result;
+}
+
 template SimpleTensor<uint16_t> depth_convert(const SimpleTensor<uint8_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
 template SimpleTensor<int16_t> depth_convert(const SimpleTensor<uint8_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
 template SimpleTensor<int32_t> depth_convert(const SimpleTensor<uint8_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
@@ -67,6 +94,8 @@ template SimpleTensor<uint8_t> depth_convert(const SimpleTensor<uint16_t> &src,
 template SimpleTensor<uint32_t> depth_convert(const SimpleTensor<uint16_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
 template SimpleTensor<uint8_t> depth_convert(const SimpleTensor<int16_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
 template SimpleTensor<int32_t> depth_convert(const SimpleTensor<int16_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
+template SimpleTensor<half> depth_convert(const SimpleTensor<float> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
+template SimpleTensor<float> depth_convert(const SimpleTensor<half> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
 } // namespace reference
 } // namespace validation
 } // namespace test
diff --git a/tests/validation/reference/DepthConvertLayer.h b/tests/validation/reference/DepthConvertLayer.h
index 1446bfda5b..5d97c73b3c 100644
--- a/tests/validation/reference/DepthConvertLayer.h
+++ b/tests/validation/reference/DepthConvertLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -35,19 +35,10 @@ namespace validation
 {
 namespace reference
 {
-template < typename T1, typename T2, typename std::enable_if < std::is_integral<T1>::value &&std::is_floating_point<T2>::value, int >::type = 0 >
-SimpleTensor<T2> depth_convert(const SimpleTensor<T1> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
-
-template < typename T1, typename T2, typename std::enable_if < std::is_floating_point<T1>::value &&std::is_integral<T2>::value, int >::type = 0 >
-SimpleTensor<T2> depth_convert(const SimpleTensor<T1> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
-
 template < typename T1, typename T2, typename std::enable_if < std::is_integral<T1>::value &&std::is_integral<T2>::value &&!std::is_same<T1, T2>::value, int >::type = 0 >
 SimpleTensor<T2> depth_convert(const SimpleTensor<T1> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
 
-template < typename T1, typename T2, typename std::enable_if < std::is_integral<T1>::value &&std::is_integral<T2>::value &&std::is_same<T1, T2>::value, int >::type = 0 >
-SimpleTensor<T2> depth_convert(const SimpleTensor<T1> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
-
-template < typename T1, typename T2, typename std::enable_if < std::is_floating_point<T1>::value &&is_floating_point<T2>::value, int >::type = 0 >
+template < typename T1, typename T2, typename std::enable_if < is_floating_point<T1>::value &&is_floating_point<T2>::value &&!std::is_same<T1, T2>::value, int >::type = 0 >
 SimpleTensor<T2> depth_convert(const SimpleTensor<T1> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
 } // namespace reference
 } // namespace validation